I made the call, set what would be hidden and what would carry over, and moved the testing programme to the MCP surface.
Two ways in
By September 2026 HYVIRAL had two ways in: an in-app agent, a chat where a manager agent routed each request to specialists next to a web video editor, and a remote MCP server, shipped in August, that let a creator’s own AI assistant call the same tools directly. MCP is the open standard that lets AI assistants such as Claude use outside tools.
Both ran on the same tool registry, credit pool and job system. The difference was who did the thinking: my agent, or theirs.
What I saw
One test round on the in-app door produced five defects and cost several engineer-days, while the same jobs succeeded over MCP at a fraction of the credits. The deeper reason was structural: general-purpose agents already hold the conversation, plan multi-step work and improve every few months without any effort from me.
What those agents cannot do is see a creator’s audience, edit real footage, render a finished video or price a paid step. The comment that opens HYVIRAL’s MCP catalogue says it plainly:
Content Intel: the proprietary-data core, and the reason an external agent connects at all. Claude and ChatGPT can generate images; they cannot see this user's audience.
On 31 August the same logic had already led me to delist ten tools: paid copies of what a host agent does natively, such as web search and image comparison.
Make MCP the product
On 24 September 2026 I wrote to the team: “I see we are wasting way too much time trying to get the in-app UI and features to work correctly, while we have the MCP surface pretty much fully operational already.”
From that day all engineering time went to the MCP surface: both testers moved to it, its defects came first, and the launch positioning became a toolkit for agents.
The tradeoff: a creator who doesn’t use an AI agent can’t edit in HYVIRAL today, and the product’s quality now depends partly on agents I don’t control.
Keep what was good, hide the rest
Most of the in-app agent was not the chat window; it was the machinery underneath, and the MCP server already ran on it. So on 2 October I hid the agent page instead of deleting it: it is unlisted, one configuration constant brings it back, and every mention left the website and the plan cards.
| Built for the in-app agent | Serving agents over MCP today |
|---|---|
| The tool registry, the one place every capability runs | The MCP bridge executes only through it; 82 tools are exposed, each with a scope and an execution contract |
| The spend confirmation before paid work | Quote before spend: a paid tool returns a price and a token, and runs only on a second call that carries it |
| Credit holds and background jobs | Jobs start with credits held; the agent polls get_job_status and can cancel |
| The planner for editing long videos | plan_edits, the same segment planner, called directly by the person’s agent |
| Quality checks on every export | The same checks return in the job’s result, with their evidence |
| The agent’s editing playbooks | The MCP skill and prompts, so any agent can follow the same sequence |
Dropped: the chat page, the manager-and-specialist planning layer, and the web editor with its export.
Results
HYVIRAL launched on 8 October 2026 as an MCP toolkit for agents. On 13 October one message produced seven posts in 19 min 12 s through the production MCP server: 455 credits quoted at most, 390 charged (credits are HYVIRAL’s prepaid units). The run is the chat on the home page.
Handing the planning to someone else’s agent moved the engineering into what that agent reads: tool descriptions, quotes written for a person, playbooks, and job results that say when to stop polling. Driving the surface as a customer found what tests missed, such as a finished job that returned no way to collect its file.